Abstract:
Aiming at the shortcomings that the active detection method of concept drift is subject to high detection delay, missed detection, and false alarm, this paper proposes an adaptive weighted concept drift detection method based on McDiarmid boundary (WMDDM). The proposed algorithm has a weighted adjustment mechanism. The adaptive attenuation algorithm is introduced as a weight function to give old data lower weights that are dynamically adjusted to adapt to the concept drift faster according to the changes in the data stream. McDiarmid's inequality is utilized to obtain the warning level and drift level of the weighted classification accuracy. While the weighted classification accuracy rate is detected to drop outside the drift level, the detection result is fed back to the classifier. While it is detected that the weighted classification accuracy rate drops beyond the warning level, the detector adapts to the change of the data flow by the triggered weight adjustment mechanism. Finally, the experiments are made on 4 artificial data sets and 1 real data set by the comparison with Fast Hoeffding Drift Detection Method (FHDDM), Drift Detection Method based on Hoeffding’s inequality (HDDM) and other algorithms. It is shown via experimental results that the proposed WMDDM algorithm has the lowest false alarm rate and missed detection rate, and its average detection delay and accuracy rate rank second among six algorithms. In addition, the proposed WMDDM algorithm is also used to classify real data sets by comparing with FHDDM algorithm, from which it is shown that WMDDM algorithm has a higher classification accuracy rate than FHDDM. Therefore, the WMDDM algorithm is more suitable for abrupt and gradual conceptual drift, and has strong robustness.